ZettaBrain RAG

Chat with your documents. Generate documents from them.

ZettaBrain plugs into NFS, SMB, S3, or OneDrive, answers questions with citations, and runs 21 Skills that turn your corpus into SOPs, quotes, compliance reports, and more. Local AI, no cloud.

Install · Try with sample data · Star the repo

Your first answer in 60 seconds

curl -fsSL https://zettabrain.app/install.sh | sudo bash

The installer sets up Python, Ollama, and the embedding model. Run the setup wizard, then launch:

sudo zettabrain-setup      # Configure storage and model
zettabrain-server          # Start web GUI at https://localhost:7860

Or install via pip:

pip install zettabrain-rag

Try with sample data

Not ready to point it at your own files? Download a test corpus and start asking questions immediately:

curl -LO https://zettabrain.io/sample-data/zettabrain-financial-test-docs.zip
unzip zettabrain-financial-test-docs.zip -d ~/zettabrain-test
zettabrain-ingest --folder ~/zettabrain-test/financial
zettabrain-chat

Ask: "What is the pre-clearance process for personal securities trades?"

You'll get an answer with citations like [Trading_Policy.docx p.3].

The problem

Before: Someone asks about the expense policy. You open the shared drive, guess which folder it's in, skim three PDFs, Ctrl+F through a 40-page handbook, and hope you found the right version.

After: You ask ZettaBrain. It searches every document you've indexed, pulls the relevant paragraphs, and gives you a sourced answer in seconds.

And when you need to write a compliance report, SOP, or RFP response? Skills pull facts from your corpus and generate a draft grounded in your actual documents, not hallucinated boilerplate.

What you get

How it works

flowchart LR
    A[Your Storage] --> B[Ingest]
    B --> C[Vector Store]
    D[Question] --> E[Hybrid Retrieval]
    C --> E
    E --> F[Local LLM]
    F --> G[Answer + Sources]

Storage: ZettaBrain reads from wherever your documents already live: a local folder, NFS share, SMB server, S3 bucket, or OneDrive.

Ingest: Documents are chunked (size tuned per file type and storage latency), embedded, and stored in ChromaDB. MD5 hashes track what's been indexed, so re-runs only process new or changed files.

Retrieval: A five-stage hybrid pipeline finds the best context for your question:

1. MMR semantic search finds chunks with similar meaning while maintaining diversity
2. BM25 keyword search catches exact terms the embedding might miss
3. Merge and deduplicate combines both result sets
4. Cross-encoder re-ranking (FlashRank) scores each chunk against your question
5. Top chunks go to the LLM with your question

Generation: The LLM sees only your question and the retrieved context. It answers based on that context and cites sources. For Skills, the same retrieval feeds a structured prompt that produces a formatted document.

Built for

Your data stays yours

ZettaBrain is private by default:

• Local inference: Ollama runs on your machine. Queries and documents never hit an external API unless you configure one.
• Local vector store: Embeddings are stored in ChromaDB on disk, not in any cloud service.
• No telemetry: No analytics, no usage tracking, no phoning home.
• Your storage: Documents stay where they are. ZettaBrain reads from your NFS/SMB/S3 mount; it doesn't copy files to a separate location.

When you enable a cloud LLM (Claude, Gemini, OpenAI), your queries and retrieved chunks are sent to that provider. The setup wizard and web UI make this choice explicit.

Skills

Skills are document generators. They combine a structured prompt with context retrieved from your corpus to produce formatted output: an executive summary, a compliance audit, an RFP response.

Built-in Skills

Price lists

Upload an Excel price list and Skills can generate quotes with real prices. Figures are tool-calculated from your data, never hallucinated by the model.

Create your own

Add a markdown file to /opt/zettabrain/src/skills/:

---
name: My Custom Skill
version: 1.0.0
description: What this skill generates
requires_corpus: true
temperature: 0.5
max_tokens: 2000
---

Your prompt instructions here. Reference {context} for retrieved documents
and {user_input} for what the user provides.

Or use the AI-assisted skill drafter in the web UI: describe what you want, and ZettaBrain generates a skill definition from your corpus.

Pick your model

Laptop (no GPU)? Start with phi4-mini. Best reasoning-per-GB on CPU.

Have a GPU? qwen2.5:14b for quality, mistral:7b for speed.

Apple Silicon? Metal acceleration is automatic. phi4-mini or llama3.1:8b work well.

Just trying it out? Use the trial mode (25 free requests, no setup).

Performance

ZettaBrain runs on CPU, flies on GPU.

Retrieval is fast everywhere (~1 second). Generation speed depends on your hardware:

For CPU-only machines, the trial mode or a cloud LLM gives instant responses while you evaluate.

Supported platforms

Reference

Full documentation for commands, configuration, diagnostics, storage backends, API endpoints, and uninstallation:

docs/REFERENCE.md

Roadmap

• Scheduled ingestion (auto-sync on interval)
• Multi-user authentication
• More built-in Skills
• Slack and Teams integration
• API rate limiting and usage tracking

Have a feature request? Open an issue.

Contributing

Contributions welcome. Please open an issue to discuss before submitting large changes.

git clone https://github.com/zettabrain/zettabrain-rag.git
cd zettabrain-rag
pip install -e ".[dev]"
pytest

License

MIT. See LICENSE.

Built by ZettaBrain
If this helps you, star the repo.